嵌入空洞卷积模块的改进YOLOv3车辆检测算法

Improved YOLOv3 Vehicle Detection Algorithm Embedded in Dilated Convolution Module

对图像或者视频中的车辆进行检测是计算机视觉领域研究的热点之一,同时也是智能交通系统的重要组成部分。鉴于车辆检测场景复杂多变以及现有的车辆检测算法不能同时满足高精度以及高实时性的要求,本文提出一种改进的YOLOv3车辆检测算法,并自建车辆检测数据集。首先在原有及特征提取网络Darknet-53中嵌入空洞卷积模块,以减少目标信息的丢失增强感受野;其次为减少错检漏检的情况,本文对传统的NMS算法进行改进,若预测框的IoU大于设定的阈值,使其以一定的方式衰减。该改进的方法在KITTI标准数据集上显示出优于其他算法的性能,同时在自建的数据集中进行验证,精度可达96%,检测速度达25.9帧/s。

Vehicle detection on image or video data is one of the hotspots in the field of computer vision,and it is also an important part of intelligent transportation systems.In view of the complex and changeable vehicle detection scenes and the existing vehicle detection algorithms can not meet the requirements of high precision and high real-time at the same time,this paper proposes an improved YOLOv3 vehicle detection algorithm and builds its own vehicle detection data set.First,we embed the dilated convolution module in the original and feature extraction network Darknet-53 to reduce the loss of target information and enhance the receptive field.Secondly,in the NMS(non-maximum suppression)module,in order to reduce the missed detection,this article discusses the traditional NMS and makes improvements.If the IoU of the prediction frame is greater than the set threshold,it will be attenuated in a certain way.The improved method shows better performance than other algorithms on the KITTI standard data set,and the verification accuracy can reach 96%in the self-built data set,and the detection speed is 25.9 frames/s.